[Aftabbs]

Summary

SparseVector accepted any NumPy array and only failed (or silently misbehaved) deep inside _from_dense. Vector, HalfVector, Bit and the scipy path of SparseVector all raise ValueError('expected ndim to be 1') for non-1D input; this makes the dense NumPy path consistent.

Root Cause / Motivation

_from_dense does len(value) and float(v) per element:

• SparseVector(np.zeros((2, 2))) raised TypeError: only length-1 arrays can be converted to Python scalars
• SparseVector(np.array(5.0)) raised TypeError: len() of unsized object
• SparseVector(np.array([[1], [0], [2]])) was silently accepted as a 3-dimensional vector (each size-1 row coerced by float()), while Vector rejects the same input

Callers that catch ValueError for bad input (e.g. the SQLAlchemy test_bad_ndim pattern) never see these cases.

Changes

File	What changed
pgvector/sparsevec.py	Raise ValueError('expected ndim to be 1') in _from_dense for non-1D ndarrays
tests/test_sparse_vector.py	test_init_ndarray_ndim covers 2D, column, and 0-d arrays (fails before the fix)
CHANGELOG.md	Entry under 0.5.2

Testing

• Existing tests pass (123 across sparse/vector/halfvec/bit)
• New test fails without the change, passes with it

Impact

Consistent, catchable error for bad NumPy input across all vector types; no behavior change for valid 1D input.

🤖 Generated with Claude Code